Self-supervised learning based anomaly detection in online social media

Abstract

Online Social Media (OSM) produce enormous data related to the human behaviours based on their interactions. One such data is the opinions expressed and posted for any specific issue addressed in the OSM. Majority of the opinions posted would be categorized as positive, negative and neutral. The lighter group's opinions are termed anomalous as it is not conforming the regular opinions posted by other users. Though, lot of conventional classification and clustering based learning algorithms works well under supervised and un-supervised environment, due to the inherent ambiguity in the tweeted data, anomaly detection poses a bigger challenge in text mining. Though the data is un-supervised, for the learning purpose it is treated as Supervised Learning by assigning class labels for the training data. This paper attempts to give an insight into various anomalies of OSM and identify behavioural anomalies for a Twitter Dataset on user's opinions on demonetization policy in India. Through Self-Supervised learning, it is observed that 86% of the user's opinions did agree to the demonetization policy and the remaining have posted negative opinions for the policy implemented.